Potential mistake

[Kalyan0821]

https://github.com/ejones313/auditing-llms/blob/04b9306f17b01df1afd8f637038928bafd654115/arca.py#L71C19-L71C19

Why are stacked_cur_toks and curr_toks_tensor here updated with the random tokens? I thought this was only required at the input, i.e., for full_embeddings in L67.